diff --git a/crates/paint/src/bg_erase.rs b/crates/paint/src/bg_erase.rs
index 751d58519b..a47c5acbfb 100644
--- a/crates/paint/src/bg_erase.rs
+++ b/crates/paint/src/bg_erase.rs
@@ -17,7 +17,11 @@ use crate::retouch::{Footprint, Region, alpha_index, apply_dab_stroke};
 pub struct BgEraseSettings {
     pub sampling: Sampling,
     pub limits: Limits,
-    /// 0..1: largest per-channel difference from the sampled colour that is erased.
+    /// 0..1: how far from the sampled colour a colour is still partly erased. The distance is the
+    /// largest increase plus the largest decrease over the colour channels (`dist`); an exact
+    /// match is erased fully, then less the further a colour is, and colours at or beyond the
+    /// Tolerance are kept (`graded_weight`). Protect Foreground Color still compares the
+    /// foreground by the largest per-channel difference (`max_channel_dist`).
     pub tolerance: f32,
     /// Keep pixels that match the foreground colour.
     pub protect_foreground: bool,
@@ -52,6 +56,12 @@ fn dist(a: &[f32], b: &[f32]) -> f32 {
     up + down
 }
 
+/// The largest per-channel difference of `a` from `b`: the distance Protect Foreground Color keeps
+/// using (#11901 measured only the sample match; #11903 keeps today's measure for the foreground).
+fn max_channel_dist(a: &[f32], b: &[f32]) -> f32 {
+    a.iter().zip(b).fold(0.0f32, |m, (x, y)| m.max((x - y).abs()))
+}
+
 /// The match weight of one dab for a colour `d` away from the sampled colour (#11901): an exact
 /// match is erased fully, then the weight falls off linearly over the whole Tolerance `tol` after
 /// a short plateau `d0 = min(tol / 20, 8 levels)`, and colours at or beyond the Tolerance are kept.
@@ -151,7 +161,7 @@ pub fn background_eraser_effect(pre: &Surface, bs: &BgEraseSettings) -> Option<i
                     continue;
                 }
                 let c = &p[..nc];
-                if protect_foreground && dist(c, &fg) <= tol {
+                if protect_foreground && max_channel_dist(c, &fg) <= tol {
                     continue;
                 }
                 weight[(y - r.y0) as usize * w + (x - r.x0) as usize] = match_w(dist(c, &sampled));
diff --git a/crates/paint/src/bg_erase_graded_alpha_tests.rs b/crates/paint/src/bg_erase_graded_alpha_tests.rs
index 51e1bbc20d..dcbfab1a20 100644
--- a/crates/paint/src/bg_erase_graded_alpha_tests.rs
+++ b/crates/paint/src/bg_erase_graded_alpha_tests.rs
@@ -175,3 +175,15 @@ fn the_distance_is_measured_in_the_documents_own_channels() {
         }
     }
 }
+
+#[test]
+fn protect_foreground_keeps_todays_distance_to_the_foreground() {
+    // #11903 measured the foreground only one channel away and keeps today's measure otherwise:
+    // the largest per-channel difference. This foreground is 100 levels off in red and in green
+    // (opposite signs): within Tolerance 50 by that measure, though the sample-match distance of
+    // #11901 adds the two to 200. The pixel is the sampled colour, so only the protection keeps it.
+    let bs = BgEraseSettings { protect_foreground: true, foreground: rgb([100, 140, 40]), ..swatch(50.0, rgb(SAMPLE)) };
+    assert_eq!(one_click(PixelFormat::RGBA8, rgb(SAMPLE), &bs), 255);
+    // Protection off: the same click erases the sample fully.
+    assert_eq!(one_click(PixelFormat::RGBA8, rgb(SAMPLE), &BgEraseSettings { protect_foreground: false, ..bs }), 0);
+}